]> git.sesse.net Git - nageru/blob - mixer.cpp
Make the last pointers in CaptureCard into unique_ptr; the amount of manual bookkeepi...
[nageru] / mixer.cpp
1 #undef Success
2
3 #include "mixer.h"
4
5 #include <assert.h>
6 #include <epoxy/egl.h>
7 #include <movit/effect_chain.h>
8 #include <movit/effect_util.h>
9 #include <movit/flat_input.h>
10 #include <movit/image_format.h>
11 #include <movit/init.h>
12 #include <movit/resource_pool.h>
13 #include <stdint.h>
14 #include <stdio.h>
15 #include <stdlib.h>
16 #include <sys/resource.h>
17 #include <algorithm>
18 #include <chrono>
19 #include <condition_variable>
20 #include <cstddef>
21 #include <cstdint>
22 #include <memory>
23 #include <mutex>
24 #include <ratio>
25 #include <string>
26 #include <thread>
27 #include <utility>
28 #include <vector>
29
30 #include "DeckLinkAPI.h"
31 #include "LinuxCOM.h"
32 #include "alsa_output.h"
33 #include "bmusb/bmusb.h"
34 #include "bmusb/fake_capture.h"
35 #include "chroma_subsampler.h"
36 #include "context.h"
37 #include "decklink_capture.h"
38 #include "decklink_output.h"
39 #include "defs.h"
40 #include "disk_space_estimator.h"
41 #include "flags.h"
42 #include "input_mapping.h"
43 #include "pbo_frame_allocator.h"
44 #include "ref_counted_gl_sync.h"
45 #include "resampling_queue.h"
46 #include "timebase.h"
47 #include "video_encoder.h"
48
49 class IDeckLink;
50 class QOpenGLContext;
51
52 using namespace movit;
53 using namespace std;
54 using namespace std::chrono;
55 using namespace std::placeholders;
56 using namespace bmusb;
57
58 Mixer *global_mixer = nullptr;
59 bool uses_mlock = false;
60
61 namespace {
62
63 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
64 {
65         if (interlaced) {
66                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
67                         input_state->buffered_frames[card_index][frame_num] =
68                                 input_state->buffered_frames[card_index][frame_num - 1];
69                 }
70                 input_state->buffered_frames[card_index][0] = { frame, field_num };
71         } else {
72                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
73                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
74                 }
75         }
76 }
77
78 }  // namespace
79
80 void QueueLengthPolicy::update_policy(int queue_length)
81 {
82         if (queue_length < 0) {  // Starvation.
83                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
84                         ++safe_queue_length;
85                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
86                                 card_index, safe_queue_length);
87                 }
88                 frames_with_at_least_one = 0;
89                 been_at_safe_point_since_last_starvation = false;
90                 return;
91         }
92         if (queue_length > 0) {
93                 if (queue_length >= int(safe_queue_length)) {
94                         been_at_safe_point_since_last_starvation = true;
95                 }
96                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
97                         --safe_queue_length;
98                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
99                                 card_index, safe_queue_length);
100                         frames_with_at_least_one = 0;
101                 }
102         } else {
103                 frames_with_at_least_one = 0;
104         }
105 }
106
107 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
108         : httpd(),
109           num_cards(num_cards),
110           mixer_surface(create_surface(format)),
111           h264_encoder_surface(create_surface(format)),
112           decklink_output_surface(create_surface(format)),
113           audio_mixer(num_cards)
114 {
115         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
116         check_error();
117
118         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
119         // will be halved when sampling them, and we need to compensate here.
120         movit_texel_subpixel_precision /= 2.0;
121
122         resource_pool.reset(new ResourcePool);
123         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
124         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
125                 output_channel[i].parent = this;
126                 output_channel[i].channel = i;
127         }
128
129         ImageFormat inout_format;
130         inout_format.color_space = COLORSPACE_sRGB;
131         inout_format.gamma_curve = GAMMA_sRGB;
132
133         // Display chain; shows the live output produced by the main chain (its RGBA version).
134         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
135         check_error();
136         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
137         display_chain->add_input(display_input);
138         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
139         display_chain->set_dither_bits(0);  // Don't bother.
140         display_chain->finalize();
141
142         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
143
144         // Start listening for clients only once VideoEncoder has written its header, if any.
145         httpd.start(9095);
146
147         // First try initializing the then PCI devices, then USB, then
148         // fill up with fake cards until we have the desired number of cards.
149         unsigned num_pci_devices = 0;
150         unsigned card_index = 0;
151
152         {
153                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
154                 if (decklink_iterator != nullptr) {
155                         for ( ; card_index < num_cards; ++card_index) {
156                                 IDeckLink *decklink;
157                                 if (decklink_iterator->Next(&decklink) != S_OK) {
158                                         break;
159                                 }
160
161                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
162                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
163                                 output->set_device(decklink);
164                                 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
165                                 ++num_pci_devices;
166                         }
167                         decklink_iterator->Release();
168                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
169                 } else {
170                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
171                 }
172         }
173
174         unsigned num_usb_devices = BMUSBCapture::num_cards();
175         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
176                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
177                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
178                 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
179         }
180         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
181
182         unsigned num_fake_cards = 0;
183         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
184                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
185                 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
186         }
187
188         if (num_fake_cards > 0) {
189                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
190         }
191
192         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
193         BMUSBCapture::start_bm_thread();
194
195         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
196                 cards[card_index].queue_length_policy.reset(card_index);
197         }
198
199         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
200
201         if (global_flags.enable_alsa_output) {
202                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
203         }
204         if (global_flags.output_card != -1) {
205                 set_output_card(global_flags.output_card);
206         }
207 }
208
209 Mixer::~Mixer()
210 {
211         BMUSBCapture::stop_bm_thread();
212
213         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
214                 {
215                         unique_lock<mutex> lock(card_mutex);
216                         cards[card_index].should_quit = true;  // Unblock thread.
217                         cards[card_index].new_frames_changed.notify_all();
218                 }
219                 cards[card_index].capture->stop_dequeue_thread();
220                 if (cards[card_index].output) {
221                         cards[card_index].output->end_output();
222                         cards[card_index].output.reset();
223                 }
224         }
225
226         video_encoder.reset(nullptr);
227 }
228
229 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
230 {
231         printf("Configuring card %d...\n", card_index);
232
233         CaptureCard *card = &cards[card_index];
234         if (card->capture != nullptr) {
235                 card->capture->stop_dequeue_thread();
236         }
237         card->capture.reset(capture);
238         card->is_fake_capture = is_fake_capture;
239         if (card->output.get() != output) {
240                 card->output.reset(output);
241         }
242         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
243         if (card->frame_allocator == nullptr) {
244                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
245         }
246         card->capture->set_video_frame_allocator(card->frame_allocator.get());
247         if (card->surface == nullptr) {
248                 card->surface = create_surface_with_same_format(mixer_surface);
249         }
250         while (!card->new_frames.empty()) card->new_frames.pop();
251         card->last_timecode = -1;
252         card->capture->configure_card();
253
254         // NOTE: start_bm_capture() happens in thread_func().
255
256         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
257         audio_mixer.reset_resampler(device);
258         audio_mixer.set_display_name(device, card->capture->get_description());
259         audio_mixer.trigger_state_changed_callback();
260 }
261
262 void Mixer::set_output_card(int card_index)
263 {
264         if (card_index == output_card_index) {
265                 return;
266         }
267         unique_lock<mutex> lock(card_mutex);
268         if (output_card_index != -1) {
269                 // Switch the old card from output to input.
270                 CaptureCard *old_card = &cards[output_card_index];
271                 old_card->output->end_output();
272
273                 old_card->capture->stop_dequeue_thread();
274                 old_card->capture = move(old_card->parked_capture);
275                 old_card->is_fake_capture = false;
276                 old_card->capture->start_bm_capture();
277         }
278
279         CaptureCard *card = &cards[card_index];
280         card->capture->stop_dequeue_thread();
281         card->parked_capture = move(card->capture);
282         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
283         configure_card(card_index, capture, /*is_fake_capture=*/true, card->output.release());
284         card->queue_length_policy.reset(card_index);
285         card->capture->start_bm_capture();
286         card->output->start_output(bmdModeHD720p5994, pts_int);  // FIXME
287         output_card_index = card_index;
288 }
289
290 namespace {
291
292 int unwrap_timecode(uint16_t current_wrapped, int last)
293 {
294         uint16_t last_wrapped = last & 0xffff;
295         if (current_wrapped > last_wrapped) {
296                 return (last & ~0xffff) | current_wrapped;
297         } else {
298                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
299         }
300 }
301
302 }  // namespace
303
304 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
305                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
306                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
307 {
308         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
309         CaptureCard *card = &cards[card_index];
310
311         if (is_mode_scanning[card_index]) {
312                 if (video_format.has_signal) {
313                         // Found a stable signal, so stop scanning.
314                         is_mode_scanning[card_index] = false;
315                 } else {
316                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
317                         steady_clock::time_point now = steady_clock::now();
318                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
319                         if (sec_since_last_switch > switch_time_s) {
320                                 // It isn't this mode; try the next one.
321                                 mode_scanlist_index[card_index]++;
322                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
323                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
324                                 last_mode_scan_change[card_index] = now;
325                         }
326                 }
327         }
328
329         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
330         assert(frame_length > 0);
331
332         size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
333         if (num_samples > OUTPUT_FREQUENCY / 10) {
334                 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
335                         card_index, int(audio_frame.len), int(audio_offset),
336                         timecode, int(video_frame.len), int(video_offset), video_format.id);
337                 if (video_frame.owner) {
338                         video_frame.owner->release_frame(video_frame);
339                 }
340                 if (audio_frame.owner) {
341                         audio_frame.owner->release_frame(audio_frame);
342                 }
343                 return;
344         }
345
346         int dropped_frames = 0;
347         if (card->last_timecode != -1) {
348                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
349         }
350
351         // Number of samples per frame if we need to insert silence.
352         // (Could be nonintegral, but resampling will save us then.)
353         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
354
355         if (dropped_frames > MAX_FPS * 2) {
356                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
357                         card_index, card->last_timecode, timecode);
358                 audio_mixer.reset_resampler(device);
359                 dropped_frames = 0;
360         } else if (dropped_frames > 0) {
361                 // Insert silence as needed.
362                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
363                         card_index, dropped_frames, timecode);
364
365                 bool success;
366                 do {
367                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
368                 } while (!success);
369         }
370
371         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
372
373         // Done with the audio, so release it.
374         if (audio_frame.owner) {
375                 audio_frame.owner->release_frame(audio_frame);
376         }
377
378         card->last_timecode = timecode;
379
380         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
381         if (video_frame.len - video_offset == 0 ||
382             video_frame.len - video_offset != expected_length) {
383                 if (video_frame.len != 0) {
384                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
385                                 card_index, video_frame.len - video_offset, expected_length);
386                 }
387                 if (video_frame.owner) {
388                         video_frame.owner->release_frame(video_frame);
389                 }
390
391                 // Still send on the information that we _had_ a frame, even though it's corrupted,
392                 // so that pts can go up accordingly.
393                 {
394                         unique_lock<mutex> lock(card_mutex);
395                         CaptureCard::NewFrame new_frame;
396                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
397                         new_frame.length = frame_length;
398                         new_frame.interlaced = false;
399                         new_frame.dropped_frames = dropped_frames;
400                         card->new_frames.push(move(new_frame));
401                         card->new_frames_changed.notify_all();
402                 }
403                 return;
404         }
405
406         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
407
408         unsigned num_fields = video_format.interlaced ? 2 : 1;
409         steady_clock::time_point frame_upload_start;
410         bool interlaced_stride = false;
411         if (video_format.interlaced) {
412                 // Send the two fields along as separate frames; the other side will need to add
413                 // a deinterlacer to actually get this right.
414                 assert(video_format.height % 2 == 0);
415                 video_format.height /= 2;
416                 assert(frame_length % 2 == 0);
417                 frame_length /= 2;
418                 num_fields = 2;
419                 if (video_format.second_field_start == 1) {
420                         interlaced_stride = true;
421                 }
422                 frame_upload_start = steady_clock::now();
423         }
424         userdata->last_interlaced = video_format.interlaced;
425         userdata->last_has_signal = video_format.has_signal;
426         userdata->last_is_connected = video_format.is_connected;
427         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
428         userdata->last_frame_rate_den = video_format.frame_rate_den;
429         RefCountedFrame frame(video_frame);
430
431         // Upload the textures.
432         size_t cbcr_width = video_format.width / 2;
433         size_t cbcr_offset = video_offset / 2;
434         size_t y_offset = video_frame.size / 2 + video_offset / 2;
435
436         for (unsigned field = 0; field < num_fields; ++field) {
437                 // Put the actual texture upload in a lambda that is executed in the main thread.
438                 // It is entirely possible to do this in the same thread (and it might even be
439                 // faster, depending on the GPU and driver), but it appears to be trickling
440                 // driver bugs very easily.
441                 //
442                 // Note that this means we must hold on to the actual frame data in <userdata>
443                 // until the upload command is run, but we hold on to <frame> much longer than that
444                 // (in fact, all the way until we no longer use the texture in rendering).
445                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
446                         unsigned field_start_line;
447                         if (field == 1) {
448                                 field_start_line = video_format.second_field_start;
449                         } else {
450                                 field_start_line = video_format.extra_lines_top;
451                         }
452
453                         if (userdata->tex_y[field] == 0 ||
454                             userdata->tex_cbcr[field] == 0 ||
455                             video_format.width != userdata->last_width[field] ||
456                             video_format.height != userdata->last_height[field]) {
457                                 // We changed resolution since last use of this texture, so we need to create
458                                 // a new object. Note that this each card has its own PBOFrameAllocator,
459                                 // we don't need to worry about these flip-flopping between resolutions.
460                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
461                                 check_error();
462                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
463                                 check_error();
464                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
465                                 check_error();
466                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
467                                 check_error();
468                                 userdata->last_width[field] = video_format.width;
469                                 userdata->last_height[field] = video_format.height;
470                         }
471
472                         GLuint pbo = userdata->pbo;
473                         check_error();
474                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
475                         check_error();
476
477                         size_t field_y_start = y_offset + video_format.width * field_start_line;
478                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
479
480                         if (global_flags.flush_pbos) {
481                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
482                                 check_error();
483                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
484                                 check_error();
485                         }
486
487                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
488                         check_error();
489                         if (interlaced_stride) {
490                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, cbcr_width * 2);
491                                 check_error();
492                         } else {
493                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
494                                 check_error();
495                         }
496                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
497                         check_error();
498                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
499                         check_error();
500                         if (interlaced_stride) {
501                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, video_format.width * 2);
502                                 check_error();
503                         } else {
504                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
505                                 check_error();
506                         }
507                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
508                         check_error();
509                         glBindTexture(GL_TEXTURE_2D, 0);
510                         check_error();
511                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
512                         check_error();
513                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
514                         check_error();
515                 };
516
517                 if (field == 1) {
518                         // Don't upload the second field as fast as we can; wait until
519                         // the field time has approximately passed. (Otherwise, we could
520                         // get timing jitter against the other sources, and possibly also
521                         // against the video display, although the latter is not as critical.)
522                         // This requires our system clock to be reasonably close to the
523                         // video clock, but that's not an unreasonable assumption.
524                         steady_clock::time_point second_field_start = frame_upload_start +
525                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
526                         this_thread::sleep_until(second_field_start);
527                 }
528
529                 {
530                         unique_lock<mutex> lock(card_mutex);
531                         CaptureCard::NewFrame new_frame;
532                         new_frame.frame = frame;
533                         new_frame.length = frame_length;
534                         new_frame.field = field;
535                         new_frame.interlaced = video_format.interlaced;
536                         new_frame.upload_func = upload_func;
537                         new_frame.dropped_frames = dropped_frames;
538                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
539                         card->new_frames.push(move(new_frame));
540                         card->new_frames_changed.notify_all();
541                 }
542         }
543 }
544
545 void Mixer::bm_hotplug_add(libusb_device *dev)
546 {
547         lock_guard<mutex> lock(hotplug_mutex);
548         hotplugged_cards.push_back(dev);
549 }
550
551 void Mixer::bm_hotplug_remove(unsigned card_index)
552 {
553         cards[card_index].new_frames_changed.notify_all();
554 }
555
556 void Mixer::thread_func()
557 {
558         eglBindAPI(EGL_OPENGL_API);
559         QOpenGLContext *context = create_context(mixer_surface);
560         if (!make_current(context, mixer_surface)) {
561                 printf("oops\n");
562                 exit(1);
563         }
564
565         // Start the actual capture. (We don't want to do it before we're actually ready
566         // to process output frames.)
567         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
568                 if (int(card_index) != output_card_index) {
569                         cards[card_index].capture->start_bm_capture();
570                 }
571         }
572
573         steady_clock::time_point start, now;
574         start = steady_clock::now();
575
576         int frame = 0;
577         int stats_dropped_frames = 0;
578
579         while (!should_quit) {
580                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
581                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
582
583                 bool master_card_is_output;
584                 unsigned master_card_index;
585                 if (output_card_index != -1) {
586                         master_card_is_output = true;
587                         master_card_index = output_card_index;
588                 } else {
589                         master_card_is_output = false;
590                         master_card_index = theme->map_signal(master_clock_channel);
591                         assert(master_card_index < num_cards);
592                 }
593
594                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
595                 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration);
596                 stats_dropped_frames += output_frame_info.dropped_frames;
597
598                 handle_hotplugged_cards();
599
600                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
601                         if (card_index == master_card_index || !has_new_frame[card_index]) {
602                                 continue;
603                         }
604                         if (new_frames[card_index].frame->len == 0) {
605                                 ++new_frames[card_index].dropped_frames;
606                         }
607                         if (new_frames[card_index].dropped_frames > 0) {
608                                 printf("Card %u dropped %d frames before this\n",
609                                         card_index, int(new_frames[card_index].dropped_frames));
610                         }
611                 }
612
613                 // If the first card is reporting a corrupted or otherwise dropped frame,
614                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
615                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
616                         ++stats_dropped_frames;
617                         pts_int += new_frames[master_card_index].length;
618                         continue;
619                 }
620
621                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
622                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
623                                 continue;
624
625                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
626                         assert(new_frame->frame != nullptr);
627                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
628                         check_error();
629
630                         // The new texture might need uploading before use.
631                         if (new_frame->upload_func) {
632                                 new_frame->upload_func();
633                                 new_frame->upload_func = nullptr;
634                         }
635                 }
636
637                 int64_t frame_duration = output_frame_info.frame_duration;
638                 render_one_frame(frame_duration);
639                 ++frame;
640                 pts_int += frame_duration;
641
642                 now = steady_clock::now();
643                 double elapsed = duration<double>(now - start).count();
644                 if (frame % 100 == 0) {
645                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
646                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
647                                 1e3 * elapsed / frame);
648                 //      chain->print_phase_timing();
649
650                         // Check our memory usage, to see if we are close to our mlockall()
651                         // limit (if at all set).
652                         rusage used;
653                         if (getrusage(RUSAGE_SELF, &used) == -1) {
654                                 perror("getrusage(RUSAGE_SELF)");
655                                 assert(false);
656                         }
657
658                         if (uses_mlock) {
659                                 rlimit limit;
660                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
661                                         perror("getrlimit(RLIMIT_MEMLOCK)");
662                                         assert(false);
663                                 }
664
665                                 if (limit.rlim_cur == 0) {
666                                         printf(", using %ld MB memory (locked)",
667                                                 long(used.ru_maxrss / 1024));
668                                 } else {
669                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
670                                                 long(used.ru_maxrss / 1024),
671                                                 long(limit.rlim_cur / 1048576),
672                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
673                                 }
674                         } else {
675                                 printf(", using %ld MB memory (not locked)",
676                                         long(used.ru_maxrss / 1024));
677                         }
678
679                         printf("\n");
680                 }
681
682
683                 if (should_cut.exchange(false)) {  // Test and clear.
684                         video_encoder->do_cut(frame);
685                 }
686
687 #if 0
688                 // Reset every 100 frames, so that local variations in frame times
689                 // (especially for the first few frames, when the shaders are
690                 // compiled etc.) don't make it hard to measure for the entire
691                 // remaining duration of the program.
692                 if (frame == 10000) {
693                         frame = 0;
694                         start = now;
695                 }
696 #endif
697                 check_error();
698         }
699
700         resource_pool->clean_context();
701 }
702
703 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
704 {
705         if (output_card_index != -1) {
706                 // The output card (ie., cards[output_card_index].output) is the master clock,
707                 // so no input card (ie., cards[card_index].capture) is.
708                 return false;
709         }
710         return (card_index == master_card_index);
711 }
712
713 Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS])
714 {
715         OutputFrameInfo output_frame_info;
716 start:
717         unique_lock<mutex> lock(card_mutex, defer_lock);
718         if (master_card_is_output) {
719                 // Clocked to the output, so wait for it to be ready for the next frame.
720                 cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration);
721                 lock.lock();
722         } else {
723                 // Wait for the master card to have a new frame.
724                 // TODO: Add a timeout.
725                 lock.lock();
726                 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
727         }
728
729         if (master_card_is_output) {
730                 handle_hotplugged_cards();
731         } else if (cards[master_card_index].new_frames.empty()) {
732                 // We were woken up, but not due to a new frame. Deal with it
733                 // and then restart.
734                 assert(cards[master_card_index].capture->get_disconnected());
735                 handle_hotplugged_cards();
736                 goto start;
737         }
738
739         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
740                 CaptureCard *card = &cards[card_index];
741                 if (card->new_frames.empty()) {
742                         assert(!input_card_is_master_clock(card_index, master_card_index));
743                         card->queue_length_policy.update_policy(-1);
744                         continue;
745                 }
746                 new_frames[card_index] = move(card->new_frames.front());
747                 has_new_frame[card_index] = true;
748                 card->new_frames.pop();
749                 card->new_frames_changed.notify_all();
750
751                 if (input_card_is_master_clock(card_index, master_card_index)) {
752                         // We don't use the queue length policy for the master card,
753                         // but we will if it stops being the master. Thus, clear out
754                         // the policy in case we switch in the future.
755                         card->queue_length_policy.reset(card_index);
756                 } else {
757                         // If we have excess frames compared to the policy for this card,
758                         // drop frames from the head.
759                         card->queue_length_policy.update_policy(card->new_frames.size());
760                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
761                                 card->new_frames.pop();
762                         }
763                 }
764         }
765
766         if (!master_card_is_output) {
767                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
768                 output_frame_info.frame_duration = new_frames[master_card_index].length;
769         }
770
771         // This might get off by a fractional sample when changing master card
772         // between ones with different frame rates, but that's fine.
773         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
774         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
775         fractional_samples = num_samples_times_timebase % TIMEBASE;
776         assert(output_frame_info.num_samples >= 0);
777
778         return output_frame_info;
779 }
780
781 void Mixer::handle_hotplugged_cards()
782 {
783         // Check for cards that have been disconnected since last frame.
784         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
785                 CaptureCard *card = &cards[card_index];
786                 if (card->capture->get_disconnected()) {
787                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
788                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
789                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
790                         card->queue_length_policy.reset(card_index);
791                         card->capture->start_bm_capture();
792                 }
793         }
794
795         // Check for cards that have been connected since last frame.
796         vector<libusb_device *> hotplugged_cards_copy;
797         {
798                 lock_guard<mutex> lock(hotplug_mutex);
799                 swap(hotplugged_cards, hotplugged_cards_copy);
800         }
801         for (libusb_device *new_dev : hotplugged_cards_copy) {
802                 // Look for a fake capture card where we can stick this in.
803                 int free_card_index = -1;
804                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
805                         if (cards[card_index].is_fake_capture) {
806                                 free_card_index = card_index;
807                                 break;
808                         }
809                 }
810
811                 if (free_card_index == -1) {
812                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
813                         libusb_unref_device(new_dev);
814                 } else {
815                         // BMUSBCapture takes ownership.
816                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
817                         CaptureCard *card = &cards[free_card_index];
818                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
819                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
820                         card->queue_length_policy.reset(free_card_index);
821                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
822                         capture->start_bm_capture();
823                 }
824         }
825 }
826
827
828 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
829 {
830         // Resample the audio as needed, including from previously dropped frames.
831         assert(num_cards > 0);
832         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
833                 const bool dropped_frame = (frame_num != dropped_frames);
834                 {
835                         // Signal to the audio thread to process this frame.
836                         // Note that if the frame is a dropped frame, we signal that
837                         // we don't want to use this frame as base for adjusting
838                         // the resampler rate. The reason for this is that the timing
839                         // of these frames is often way too late; they typically don't
840                         // “arrive” before we synthesize them. Thus, we could end up
841                         // in a situation where we have inserted e.g. five audio frames
842                         // into the queue before we then start pulling five of them
843                         // back out. This makes ResamplingQueue overestimate the delay,
844                         // causing undue resampler changes. (We _do_ use the last,
845                         // non-dropped frame; perhaps we should just discard that as well,
846                         // since dropped frames are expected to be rare, and it might be
847                         // better to just wait until we have a slightly more normal situation).
848                         unique_lock<mutex> lock(audio_mutex);
849                         bool adjust_rate = !dropped_frame;
850                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
851                         audio_task_queue_changed.notify_one();
852                 }
853                 if (dropped_frame) {
854                         // For dropped frames, increase the pts. Note that if the format changed
855                         // in the meantime, we have no way of detecting that; we just have to
856                         // assume the frame length is always the same.
857                         pts_int += length_per_frame;
858                 }
859         }
860 }
861
862 void Mixer::render_one_frame(int64_t duration)
863 {
864         // Get the main chain from the theme, and set its state immediately.
865         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
866         EffectChain *chain = theme_main_chain.chain;
867         theme_main_chain.setup_chain();
868         //theme_main_chain.chain->enable_phase_timing(true);
869
870         GLuint y_tex, cbcr_tex;
871         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
872         assert(got_frame);
873
874         // Render main chain.
875         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
876         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height);  // Saves texture bandwidth, although dithering gets messed up.
877         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
878         check_error();
879         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
880         resource_pool->release_fbo(fbo);
881
882         chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
883         if (output_card_index != -1) {
884                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, theme_main_chain.input_frames, pts_int, duration);
885         }
886         resource_pool->release_2d_texture(cbcr_full_tex);
887
888         // Set the right state for rgba_tex.
889         glBindFramebuffer(GL_FRAMEBUFFER, 0);
890         glBindTexture(GL_TEXTURE_2D, rgba_tex);
891         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
892         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
893         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
894
895         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
896         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
897
898         // The live frame just shows the RGBA texture we just rendered.
899         // It owns rgba_tex now.
900         DisplayFrame live_frame;
901         live_frame.chain = display_chain.get();
902         live_frame.setup_chain = [this, rgba_tex]{
903                 display_input->set_texture_num(rgba_tex);
904         };
905         live_frame.ready_fence = fence;
906         live_frame.input_frames = {};
907         live_frame.temp_textures = { rgba_tex };
908         output_channel[OUTPUT_LIVE].output_frame(live_frame);
909
910         // Set up preview and any additional channels.
911         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
912                 DisplayFrame display_frame;
913                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
914                 display_frame.chain = chain.chain;
915                 display_frame.setup_chain = chain.setup_chain;
916                 display_frame.ready_fence = fence;
917                 display_frame.input_frames = chain.input_frames;
918                 display_frame.temp_textures = {};
919                 output_channel[i].output_frame(display_frame);
920         }
921 }
922
923 void Mixer::audio_thread_func()
924 {
925         while (!should_quit) {
926                 AudioTask task;
927
928                 {
929                         unique_lock<mutex> lock(audio_mutex);
930                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
931                         if (should_quit) {
932                                 return;
933                         }
934                         task = audio_task_queue.front();
935                         audio_task_queue.pop();
936                 }
937
938                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
939                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
940                 vector<float> samples_out = audio_mixer.get_output(
941                         double(task.pts_int) / TIMEBASE,
942                         task.num_samples,
943                         rate_adjustment_policy);
944
945                 // Send the samples to the sound card, then add them to the output.
946                 if (alsa) {
947                         alsa->write(samples_out);
948                 }
949                 if (output_card_index != -1) {
950                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
951                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
952                 }
953                 video_encoder->add_audio(task.pts_int, move(samples_out));
954         }
955 }
956
957 void Mixer::release_display_frame(DisplayFrame *frame)
958 {
959         for (GLuint texnum : frame->temp_textures) {
960                 resource_pool->release_2d_texture(texnum);
961         }
962         frame->temp_textures.clear();
963         frame->ready_fence.reset();
964         frame->input_frames.clear();
965 }
966
967 void Mixer::start()
968 {
969         mixer_thread = thread(&Mixer::thread_func, this);
970         audio_thread = thread(&Mixer::audio_thread_func, this);
971 }
972
973 void Mixer::quit()
974 {
975         should_quit = true;
976         audio_task_queue_changed.notify_one();
977         mixer_thread.join();
978         audio_thread.join();
979 }
980
981 void Mixer::transition_clicked(int transition_num)
982 {
983         theme->transition_clicked(transition_num, pts());
984 }
985
986 void Mixer::channel_clicked(int preview_num)
987 {
988         theme->channel_clicked(preview_num);
989 }
990
991 void Mixer::start_mode_scanning(unsigned card_index)
992 {
993         assert(card_index < num_cards);
994         if (is_mode_scanning[card_index]) {
995                 return;
996         }
997         is_mode_scanning[card_index] = true;
998         mode_scanlist[card_index].clear();
999         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1000                 mode_scanlist[card_index].push_back(mode.first);
1001         }
1002         assert(!mode_scanlist[card_index].empty());
1003         mode_scanlist_index[card_index] = 0;
1004         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1005         last_mode_scan_change[card_index] = steady_clock::now();
1006 }
1007
1008 Mixer::OutputChannel::~OutputChannel()
1009 {
1010         if (has_current_frame) {
1011                 parent->release_display_frame(&current_frame);
1012         }
1013         if (has_ready_frame) {
1014                 parent->release_display_frame(&ready_frame);
1015         }
1016 }
1017
1018 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1019 {
1020         // Store this frame for display. Remove the ready frame if any
1021         // (it was seemingly never used).
1022         {
1023                 unique_lock<mutex> lock(frame_mutex);
1024                 if (has_ready_frame) {
1025                         parent->release_display_frame(&ready_frame);
1026                 }
1027                 ready_frame = frame;
1028                 has_ready_frame = true;
1029         }
1030
1031         if (new_frame_ready_callback) {
1032                 new_frame_ready_callback();
1033         }
1034
1035         // Reduce the number of callbacks by filtering duplicates. The reason
1036         // why we bother doing this is that Qt seemingly can get into a state
1037         // where its builds up an essentially unbounded queue of signals,
1038         // consuming more and more memory, and there's no good way of collapsing
1039         // user-defined signals or limiting the length of the queue.
1040         if (transition_names_updated_callback) {
1041                 vector<string> transition_names = global_mixer->get_transition_names();
1042                 bool changed = false;
1043                 if (transition_names.size() != last_transition_names.size()) {
1044                         changed = true;
1045                 } else {
1046                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1047                                 if (transition_names[i] != last_transition_names[i]) {
1048                                         changed = true;
1049                                         break;
1050                                 }
1051                         }
1052                 }
1053                 if (changed) {
1054                         transition_names_updated_callback(transition_names);
1055                         last_transition_names = transition_names;
1056                 }
1057         }
1058         if (name_updated_callback) {
1059                 string name = global_mixer->get_channel_name(channel);
1060                 if (name != last_name) {
1061                         name_updated_callback(name);
1062                         last_name = name;
1063                 }
1064         }
1065         if (color_updated_callback) {
1066                 string color = global_mixer->get_channel_color(channel);
1067                 if (color != last_color) {
1068                         color_updated_callback(color);
1069                         last_color = color;
1070                 }
1071         }
1072 }
1073
1074 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1075 {
1076         unique_lock<mutex> lock(frame_mutex);
1077         if (!has_current_frame && !has_ready_frame) {
1078                 return false;
1079         }
1080
1081         if (has_current_frame && has_ready_frame) {
1082                 // We have a new ready frame. Toss the current one.
1083                 parent->release_display_frame(&current_frame);
1084                 has_current_frame = false;
1085         }
1086         if (has_ready_frame) {
1087                 assert(!has_current_frame);
1088                 current_frame = ready_frame;
1089                 ready_frame.ready_fence.reset();  // Drop the refcount.
1090                 ready_frame.input_frames.clear();  // Drop the refcounts.
1091                 has_current_frame = true;
1092                 has_ready_frame = false;
1093         }
1094
1095         *frame = current_frame;
1096         return true;
1097 }
1098
1099 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1100 {
1101         new_frame_ready_callback = callback;
1102 }
1103
1104 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1105 {
1106         transition_names_updated_callback = callback;
1107 }
1108
1109 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1110 {
1111         name_updated_callback = callback;
1112 }
1113
1114 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1115 {
1116         color_updated_callback = callback;
1117 }
1118
1119 mutex RefCountedGLsync::fence_lock;